Generalizability of polygenic prediction models: how is the R² defined on test data?

Area of Science:

• Genetics
• Biostatistics
• Personalized Medicine

Background:

• Polygenic risk scores (PRS) are vital for personalized medicine, quantifying genetic predisposition.
• A key challenge is ensuring PRS model generalizability across diverse ancestries.
• R-squared is commonly used to assess PRS prediction accuracy on test data.

Purpose of the Study:

• To compare three R-squared definitions for evaluating PRS generalizability on test data.
• To analyze the impact of different R-squared definitions on PRS model performance across ancestries.

Main Methods:

• Utilized UK Biobank genotype data for analysis.
• Derived polygenic models for height, BMI, and lipoprotein A using European ancestry training data.
• Evaluated models on diverse test populations, comparing three R-squared definitions.

Main Results:

• The choice of R-squared definition significantly alters PRS generalizability results.
• Correlation-based R-squared measures discriminative performance (0-1 range).
• MSPE-based R-squared assesses discrimination and calibration, potentially yielding negative values for miscalibrated predictions.

Conclusions:

• Explicitly stating the R-squared definition used is essential for reporting PRS generalizability.
• Understanding R-squared definitions aids accurate interpretation of PRS performance across populations.
• Further research is needed for well-calibrated PRS models in diverse populations.